A new Linux memory compression approach called CRAM, proposed by Meta's Gregory Price at the Linux Plumbers Conference, takes a different path from existing options like zswap and ZRAM. Instead of treating compressed memory as a swap-layer feature, CRAM keeps the compressed data in RAM and presents it as a private NUMA node. This lets Linux manage it with normal memory semantics, including migration and ballooning, while avoiding the fault and swap overhead that dominates compressed-memory performance.
The performance claims are striking. In read-only workloads, CRAM was measured at 489 million operations per second versus ZRAM's 1.1 million, a roughly 452x speedup. Writes are more expensive because data must be faulted and migrated back to the original NUMA domain before modification; even so, a workload with 20% writes still ran 5.4x faster than ZRAM. A mechanism called the "Chicken Bit" tells Linux to stop using CRAM while it handles allocation pressure, preventing cascading failures the presentation colorfully calls a "poison storm."
CRAM is not fully solved. Because compressibility varies wildly, it remains unclear how to determine how much logical RAM is available or when it will run out. The source article notes that this is an ongoing research area. While the work is aimed at large Linux servers, ZRAM and zswap are widely used on consumer devices such as the Steam Deck, so CRAM could eventually bring speedups there too—if the remaining implementation questions are resolved. The report is based on conference slides rather than a live demo, so some details may still be incomplete.